Dataset Viewer
Auto-converted to Parquet Duplicate
repo_id
string
root
string
num_frames
int64
num_episodes
int64
fps
int64
features
dict
errors
list
warnings
list
strict_certificate_count
int64
certified_successes_per_task
dict
sampled_frames
int64
black_image_samples
int64
non_finite_samples
int64
sample_tasks
list
joint_names_expected
list
visiobot/radeonvla-reflex-physical-2k
/workspace/RadeonVLA-Reflex/datasets/radeonvla_reflex_physical_2k
468,889
2,000
20
{ "observation.state": { "dtype": "float32", "shape": [ 9 ] }, "action": { "dtype": "float32", "shape": [ 9 ] }, "observation.images.world": { "dtype": "video", "shape": [ 240, 320, 3 ] }, "observation.images.wrist": { "dtype": "video",...
[]
[]
2,000
{ "apple_blue_left": 100, "apple_blue_right": 100, "apple_white_left": 100, "apple_white_right": 100, "banana_blue_left": 100, "banana_blue_right": 100, "banana_white_left": 100, "banana_white_right": 100, "lemon_blue_left": 100, "lemon_blue_right": 100, "lemon_white_left": 100, "lemon_white_rig...
1,000
0
0
[ "Pick the apple and place it in the blue bowl on the left.", "Pick the apple and place it in the blue bowl on the right.", "Pick the apple and place it in the white bowl on the left.", "Pick the apple and place it in the white bowl on the right.", "Pick the banana and place it in the blue bowl on the left."...
[ "panda_joint1", "panda_joint2", "panda_joint3", "panda_joint4", "panda_joint5", "panda_joint6", "panda_joint7", "panda_finger_joint1", "panda_finger_joint2" ]

RadeonVLA-Reflex Physical-2K

Physical-2K contains 2,000 strictly validated successful Genesis episodes for language-conditioned Franka fruit sorting. Coverage is exactly five fruits × four bowl positions × 100 episodes = 2,000 episodes.

Verified release facts

Item Value
Episodes 2,000
Frames 468,889
Registered task variations 20
Episodes per task variation 100
Control frequency 20 Hz
Strict success certificates 2,000
Sampled image frames 1,000
Black sampled frames 0
Non-finite sampled frames 0
Validation errors / warnings 0 / 0

The complete machine-readable report is available at validation/dataset_validation.json (SHA256 fb1a1a0be5c8e5efa24342eeb62f962ae5a6b4b0e3465b83a86eb2460516e113).

Task coverage and language

The 20 task variations are the Cartesian product of:

  • fruit: apple, banana, lemon, orange, plum;
  • target: blue-left, blue-right, white-left, white-right.

Every task variation has exactly 100 successful episodes. The metadata contains 40 collected instruction strings because each task has two recorded language forms, for example Pick the apple and place it in the blue bowl on the left. and Sort the apple into the blue bowl on the left.

Observation and action schema

Field Shape/type Description
observation.images.world 240×320 RGB video World camera
observation.images.wrist 240×320 RGB video Wrist camera
observation.state 9-D float32 Franka arm and gripper qpos
action 9-D float32 Absolute joint-position target

Joint order is panda_joint1..7, panda_finger_joint1..2. Videos are AV1, 20 FPS, without audio. The dataset follows LeRobot codebase format v3.0.

Physical-success policy

The published manifest records strict_physics=true, kinematic_grasp_assist=false, placement_nudge=false, and a rigid-body pose write guard. Failed collection attempts were discarded. Each committed episode has an atomic certificate under certificates/ and passed schema, coverage, finite-value, image-statistics, intervention, and certificate correspondence checks.

Physical-2K combines two independently audited 1K sources generated from source commits dbc12b69e88be787b679bb41cf2026ca43e0eda6 and cbf07f9306cd175e82d1e60d1c02cc0b07d2f9d2. The immutable aggregation details remain in recording_manifest.json.

Limitations and license

This is simulation-only data for the registered fruit/bowl scene; it does not establish real-robot performance. The dataset is released under CC BY 4.0. YCB object assets retain their attribution requirements.

Downloads last month
578